I ship an on-chain AI agent at a hackathon almost every week.
Solo · building in public · all of it with Claude Code
Everyone's racing to let AI agents execute — hold the keys, move the funds, make the call. I keep building the opposite. The model can price, score, phrase, or propose — but a contract (or a hard gate) verifies and settles. The model is the optimizer; the contract is the safety boundary.
It shows up everywhere:
- Credo — an AI underwrites an under-collateralized loan, but a deterministic model sets the score and the contract enforces every cap on-chain. HashKey Chain
- Backstop — an agent posts a price; a Uniswap v4 hook decides the fee per swap, so no one can front-run it. X Layer mainnet
- Aegis — an on-chain AI committee rules if an insurance claim happened; deterministic math sets how much. Somnia
- Mandate — one sentence becomes a tree of on-chain permissions; pull the root and every agent goes dark in one tx. MetaMask Smart Accounts
Same shape outside web3, too: Steward lets an LLM hunt waste in a real cloud account, but a policy gate blocks anything destructive. Sentinel is an agent that red-teams other agents before they reach prod.
Agents that hire and pay each other, with real settlement. CAProxy takes one brief, hires a team of agents, pays each in USDC on Base mainnet, and hands back the receipts. Custodian manages a tokenized cargo at sea and pays per data call (x402) before it's allowed to act. Bulwark is the spending rails so an agent with a wallet can't be drained.
Not everything is an agent. Vouch proves an exchange is solvent in zero-knowledge without revealing a single balance. Tongue dubs a live voice room so everyone hears everyone in their own language. Plumb makes architecture rules executable and catches drift on every merge.
Solidity · Foundry · TypeScript · viem / wagmi · Next.js · ERC-4337 — and whatever the hackathon's chain needs. Every line built with Claude Code.
The full run is in the repos tab below ↓
Building in public — follow along on X.